TDD-UNet:Transformer with double decoder UNet for COVID-19 lesions segmentation

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Healthcare
• Radiology

Background:

• The COVID-19 pandemic poses significant global health risks.
• Accurate segmentation of COVID-19 lesions in medical images is crucial for diagnosis and treatment.
• Traditional UNet models struggle with long-range dependencies in images, limiting segmentation accuracy.

Purpose of the Study:

• To develop an advanced segmentation model for COVID-19 lesions.
• To address the limitations of existing UNet and Transformer models in capturing both global context and fine-grained details.

Main Methods:

• Proposed TDD-UNet, a novel architecture combining Transformer's multi-head self-attention with a dual decoder UNet.
• Integrated Transformer Encoder into the UNet encoding layer for global context extraction.
• Employed a dual decoder structure with background prediction and deep supervision for enhanced foreground segmentation.

Main Results:

• TDD-UNet demonstrated superior performance in segmenting COVID-19 lesions across CT and CXR modalities.
• Achieved higher Dice and Jaccard scores with lower standard deviation compared to existing methods.
• Ablation studies confirmed the effectiveness of key components in the TDD-UNet model.

Conclusions:

• The proposed TDD-UNet model offers improved accuracy and robustness for COVID-19 lesion segmentation.
• This approach effectively combines global context and local details for precise medical image segmentation.
• TDD-UNet shows significant potential for clinical application in COVID-19 diagnosis.